Public Blog/Website
Hi, I’m Avery. I’m a longtime product engineer turned AI researcher. I’m a Research Assistant at the National Deep Inference Fabric (NDIF) with David Bau, a member of MIT AI Alignment (MAIA), and pursuing my MS in Computer Science at Northeastern University.
Before transitioning to AI research, I spent over 12 years as a software engineer, with my longest tenure at Pivotal Labs.
I believe that understanding how AI systems learn, reason, and fail is essential for building technology we can trust. My recent work focuses on methodologies that reveal adversarial behaviors under messy, real-life situations.
Current projects explore deception detection in multi-agent games, training small reasoning models, and understanding refusal and alignment patterns in frontier open and closed models.
Multi-Agent Social Deception Arena (project lead): Evaluation platform for testing strategic deception and persuasion capabilities in frontier LLMs using a well-loved social deduction game. Live leaderboard continuously benchmarks ChatGPT, Claude, Gemini, DeepSeek, Kimi, and other models.
Showing Our Work: Training Gemma3 1B To Reason: Independent project to develop an efficient, near-SotA post-training recipe for 1B language models under limited computation budgets (under judgement) notebook here
National Deep Inference Fabric: The NSF National Deep Inference Fabric (NDIF) is a research computing project that enables researchers and students to perform mechanistic interpretability research on models, with sizes up to a 405B parameter open-weight model.
I’ve been playing classical cello since I was about 7, having previously subbed with the Boston Philharmonic, and continue to play today as part of the Mercury Orchestra and various other groups.
Feel free to find me on LinkedIn.